deepgeek/CosyVoice
2
0
Fork 0
mirror of https://github.com/FunAudioLLM/CosyVoice.git synced 2026-02-04 09:29:25 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix bug

Browse Source
This commit is contained in:
lyuxiang.lx
2025-04-16 13:48:13 +08:00
parent 2c193781cc
commit 3770c1c8b1
1 changed files with 6 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
cosyvoice/flow/decoder.py
View File

@@ -158,9 +158,12 @@ class CausalAttnProcessor2_0(AttnProcessor2_0):
key_cache = attn.to_k(encoder_hidden_states)
value_cache = attn.to_v(encoder_hidden_states)
# NOTE always concat cache for interface compatibility
key = torch.concat([cache[:, :, :, 0], key_cache], dim=1)
value = torch.concat([cache[:, :, :, 1], value_cache], dim=1)
# NOTE here we judge cache.size(0) instead of cache.size(1), because init_cache has size (2, 0, 512, 2)
if cache.size(0) != 0:
key = torch.concat([cache[:, :, :, 0], key_cache], dim=1)
value = torch.concat([cache[:, :, :, 1], value_cache], dim=1)
else:
key, value = key_cache, value_cache
cache = torch.stack([key_cache, value_cache], dim=3)
inner_dim = key.shape[-1]